Deep Learning Inter-city Road Conditions in East Africa Focusing on Rwanda for Infrastructure Prioritization using Satellite Imagery and Mobile Data

Abstract

Traditional survey methods for gathering information, such as questionnaires and field visits, have long been used in East Africa to evaluate road conditions and prioritize their development. These surveys are time-consuming, expensive, and vulnerable to human error. Road building and maintenance, on the other hand, has long experienced multiple challenges due to a lack of accountability and validation of conventional approaches to determining which areas to prioritize. With the digital revolution, a lot of data is generated daily such as call detail record (CDR), that is likely to contain useful proxy data for spatial mobility distribution across different routes. In this research we focus on satellite imagery data with applications in East Africa and Google Maps suggested inter-city roads to assess road conditions and provide an approach for infrastructure prioritization given mobility patterns between cities. With increased urban population, East African cities have been expanding in multiple directions affecting the overall distribution of residential areas and consequently likely to impact the mobility trends across cities. We introduce a novel approach for infrastructure prioritization using deep learning and big data analytics. We apply deep learning to satellite imagery, to assess road conditions by area and big data analytics to CDR data, to rank which ones could be prioritized for construction given mobility trends. Among deep learning models considered for roads condition classification, EfficientNet-B3 outperforms them and achieves accuracy of 99%.